Between facets and domains 10 aspects of the big five

Peterson University of Toronto Factor analyses of 75 facet scales from 2 major Big Five inventories, in the Eugene-Springfield community sample N ⫽ 481, produced a 2-factor solution for

Between Facets and Domains: 10 Aspects of the Big Five

Colin G DeYoung

Centre for Addiction and Mental Health

Jordan B Peterson University of Toronto

Factor analyses of 75 facet scales from 2 major Big Five inventories, in the Eugene-Springfield

community sample (N ⫽ 481), produced a 2-factor solution for the 15 facets in each domain These

findings indicate the existence of 2 distinct (but correlated) aspects within each of the Big Five, representing an intermediate level of personality structure between facets and domains The authors characterized these factors in detail at the item level by correlating factor scores with the International Personality Item Pool (L R Goldberg, 1999) These correlations allowed the construction of a 100-item measure of the 10 factors (the Big Five Aspect Scales [BFAS]), which was validated in a 2nd sample

(N ⫽ 480) Finally, the authors examined the correlations of the 10 factors with scores derived from 10

genetic factors that a previous study identified underlying the shared variance among the Revised NEO Personality Inventory facets (K L Jang et al., 2002) The correspondence was strong enough to suggest that the 10 aspects of the Big Five may have distinct biological substrates

Keywords:personality, Big Five, five factor model, aspects, facets

Personality trait dimensions can be categorized by arranging

them into hierarchies, based on their intercorrelations Broad

do-mains (e.g., Extraversion), each encompassing many related traits,

are located near the top of the hierarchy, and very specific patterns

of behavior and experience (e.g., talking a lot) are located near the

bottom The arrangement of these hierarchies has been a central

preoccupation of personality psychologists for the better part of a

century Considerable progress has been made, leading to a

rea-sonable degree of consensus regarding the makeup of an adequate

categorization scheme The five-factor model, or Big Five, which

originated from studies of trait-descriptive adjectives drawn from

the lexicon, is the most widely used classification system for

personality traits, identifying five broad domains of personality:

Extraversion, Agreeableness, Conscientiousness, Neuroticism, and

Openness/Intellect (Costa & McCrae, 1992a; Digman, 1990;

Gold-berg, 1993; John & Srivastava, 1999) Like any dominant

para-digm, the Big Five model has drawn its fair share of criticisms and

proposals for alternatives (e.g., Ashton et al., 2004; Saucier, 2003;

Waller, 1999; Zuckerman, Kuhlman, Joireman, Teta, & Kraft,

1993) Nonetheless, the Big Five has proved extremely useful in providing a common language for researchers and organizing personality research

Much research on the Big Five has focused on a two-level hierarchy, with the five domains at the top subsuming narrower traits called “facets” at a second level This approach is exempli-fied by the widely used Revised NEO Personality Inventory (NEO-PI-R; Costa & McCrae, 1992b), which breaks each of the five domains down into six facets.1More than two levels can be identified, however Since the discovery by Digman (1997) that the regular pattern of correlations among the Big Five has a higher order factor solution, there has been increasing discussion of levels

of the hierarchy above the Big Five domains (DeYoung, 2006; DeYoung, Peterson, & Higgins, 2002; Jang et al., 2006; Markon, Krueger, & Watson, 2005; Saucier, 2003) Two constructs, labeled

Alpha and Beta (Digman, 1997), or Stability and Plasticity

(DeYoung, 2006; DeYoung et al., 2002), appear to constitute the highest level of personality organization in the hierarchy built around the Big Five and have been described as “metatraits.” Less attention has been paid to a level of trait organization located between facets and domains Reasons exist, however, to suspect that this level might be both interesting and important

A behavior genetic study in large Canadian and German sam-ples found that two genetic factors are responsible for the shared variance of the six facet scales that make up each of the Big Five

in the NEO-PI-R (Jang, Livesley, Angleitner, Riemann, & Vernon,

1One might well argue that this approach includes three levels, as the items that make up each facet scale typically describe multiple distinguish-able patterns of behavior and experience (Digman, 1990) Most research linking personality ratings to other phenomena does not investigate indi-vidual items, however, for psychometric reasons

Colin G DeYoung, Department of Psychology, Yale University; Lena

C Quilty, Clinical Research Department, Centre for Addiction and Mental

Health, Toronto, Ontario, Canada; Jordan B Peterson, Department of

Psychology, University of Toronto, Toronto, Ontario, Canada

This study was supported in part by a grant from the Social Sciences and

Humanities Research Council of Canada awarded to Jordan B Peterson

We thank Weronika Sroczynski for her help in running this study, Lewis

R Goldberg for his generosity in making data available from the

Eugene-Springfield community sample, Brian P O’Connor for advice on factor

analysis, and Kerry L Jang for advice on calculating genetic factor scores

Correspondence concerning this article should be addressed to Colin G

DeYoung, Department of Psychology, Yale University, Box 208205, New

Haven, CT 06520 E-mail: cdeyoung@post.harvard.edu

880

2002) Each of the Big Five domains, therefore, appears potentially

divisible into two subdomains with distinct biological sources

This finding would, by itself, be sufficient to motivate

investiga-tion into an intermediate level of personality structure Addiinvestiga-tional

sources of motivation can be found in the personality literature,

where the possibility that one or more of the Big Five might

subsume two separable subdomains has been raised in a variety of

contexts

Depue and Collins (1999) reviewed the literature on

Extraver-sion, for example, and noted a primary division within the domain,

between agency (“social dominance and the enjoyment of

leader-ship roles, assertiveness, exhibitionism, and a subjective sense of

potency in accomplishing goals,” p 492) and sociability (They

note a third traditional conception of Extraversion as impulsivity

but argue that impulsivity is in fact a compound trait combining

Extraversion with low Conscientiousness or Constraint.) Some

empirical support for such a division can be found in factor

analyses of the NEO Personality Inventory (NEO-PI; McCrae &

Costa, 1985, which predated the NEO-PI-R and did not include

facet scales for Agreeableness and Conscientiousness) These

anal-yses demonstrated that the Assertiveness and Activity facets of

Extraversion split off in a separate factor from the other four

Extraversion facets (Church, 1994; Church & Burke, 1994) At

least one widely used instrument loosely based on the Big Five, the

Hogan Personality Inventory, reflects this division, dividing the

assessment of Extraversion between “Ambition” and “Sociability”

scales (Hogan & Hogan, 1992)

Costa, McCrae, and Dye (1991) described Conscientiousness

“as having both proactive and inhibitive aspects” (p 887), the

proactive aspect including such traits as “need for achievement and

commitment to work,” and the inhibitive aspect including such

traits as “moral scrupulousness and cautiousness.” Empirical

sup-port for a similar division is offered by a study that performed

factor analysis of scales from seven major personality inventories,

including only scales identified by their authors as conceptually

related to Conscientiousness (Roberts, Chernyshenko, Stark, &

Goldberg, 2005) Two of these instruments, the NEO-PI-R and the

Abridged Big Five Circumplex scales from the International

Per-sonality Item Pool (AB5C-IPIP; Goldberg, 1999) were specifically

designed to assess facets of the Big Five Although Roberts et al

found six factors in total, all but two of the NEO and AB5C facets

were subsumed within two factors, labeled Industriousness and

Order, suggesting that, at least as defined in Big Five space,

Conscientiousness has two primary subdomains This finding is

similar to that of Jackson, Paunonen, Fraboni, and Goffin (1996),

who found that a factor solution splitting Conscientiousness into

Achievement and Methodicalness was better than the standard Big

Five solution in their instrument, the Personality Research Form

In relation to Agreeableness, Ashton and Lee (2005) have

re-cently noted that two facets of Agreeableness in the NEO-PI-R,

Straightforwardness and Modesty, have relatively weak loadings

on Agreeableness They demonstrated that these two facets were

good markers of a factor labeled Honesty-Humility, in their

six-factor model presented as an alternative to the Big Five This

finding suggests the possibility that, within the Big Five,

Agree-ableness might be separable into two subdomains Perhaps, rather

than adding a sixth domain, as Ashton and Lee (2005; Ashton et

al., 2004) suggest, one could instead discriminate between two

aspects of Agreeableness at a level of personality organization between facets and domains

Some of the most intense debate on the Big Five has centered on how best to characterize the fifth factor, commonly labeled either

Openness to Experience or Intellect The compound label Open-ness/Intellecthas become increasingly popular precisely because both labels apparently identify distinct but equally important as-pects of the domain (DeYoung, Peterson, & Higgins, 2005; John-son, 1994; Saucier, 1992) Johnson (1994) noted that two of the purest representations of the Openness/Intellect domain, from a factoring standpoint, are the Ideas and Aesthetics facets of the NEO-PI-R These were characterized elegantly by Johnson as representing interests in truth and beauty, respectively, which may begin to capture the conceptual distinction between Intellect and Openness

Less attention has been paid to the presence of different subdo-mains within Neuroticism In reviewing lexical studies of person-ality structure, however, Saucier and Goldberg (2001) identified anxiety/fearfulness and irritability as distinct trait clusters and indicated that irritability does not always fall unambiguously within the Neuroticism factor, though it is included within Neu-roticism in the NEO-PI-R’s Angry-Hostility facet

Jang et al.’s (2002) finding that two genetic factors underlie the shared variance of the facets in each of the Big Five suggests that the trend toward identifying exactly two subfactors within each of the Big Five may represent more than mere coincidence or desire for parsimony The purpose of the present study was to extend the investigation of this level of organization within the Big Five by addressing some of the limitations of Jang et al.’s study Most important is the necessity of analyzing a reasonably comprehen-sive selection of facets within each of the Big Five domains Jang

et al examined the covariance of the six facets within each domain

of the NEO-PI-R, but the facet structure of the NEO-PI-R was derived theoretically, based on a review of the literature (Costa & McCrae, 1992b), and nothing guarantees that its facets sample the space within each domain thoroughly In addition to the NEO-PI-R, therefore, we used another instrument in the present study, the AB5C-IPIP (Goldberg, 1999), whose facet level structure was devised by an algorithm that provided more thorough coverage of the universe of personality descriptors

The AB5C-IPIP facets were derived from the AB5C lexical model developed by Hofstee, de Raad, and Goldberg (1992) The AB5C model takes advantage of the fact that almost all trait-descriptive adjectives can be represented as a blend of two Big Five dimensions Each of the 10 possible pairs of Big Five dimen-sions can therefore be used to define a circumplex, or circular arrangement of traits, with Big Five axes at 0° and 90° Facets were defined by dividing each of these 10 circumplexes with six axes, located at 15°, 45°, 75°, etc., thus defining 12 sections of 30° each Adjectives falling within each section or its polar opposite represent a facet There are two “factor-pure” facets in each

circumplex, spanning the x- and y-axes, plus four facets that

represent a positive primary loading on one of the Big Five and a positive or negative secondary loading on the other Across all 10 circumplexes, 9 facets are thus defined for each of the Big Five domains—1 factor-pure and 8 with secondary loadings Each of the AB5C-IPIP facets targeted the content of the adjectives in one

of the AB5C lexical facets, using short descriptive phrases, which are more consistently interpreted than single adjectives (Goldberg,

1999) The AB5C-IPIP provides the most thorough facet-level

coverage of the Big Five of any instrument presently available

Study 1 reports the factor analysis of facets within each Big Five

domain Study 2 uses the IPIP to characterize the resulting factors

at the item level and to provide an instrument for assessing them

Study 3 examines how similar these phenotypic factors are to the

genetic factors reported by Jang et al (2002)

Study 1

We investigated the number of factors present within the facets

of two major Big Five personality questionnaires, which provided

a total of 15 facets for each domain The NEO-PI-R was used

because it is the most widely used measure of the Big Five and it

facilitated comparisons with Jang et al.’s (2002) genetic findings

The AB5C-IPIP was used to achieve more thorough coverage of

facet-level traits than would be provided by the NEO-PI-R alone

Our hypothesis was that the most likely result for each domain was

a two-factor solution

Method

Participants. Participants were 481 members of the

Eugene-Springfield community sample (ESCS; 200 men and 281 women),

ranging in age from 20 to 85 years (M ⫽ 52.51, SD ⫽ 12.63), who

completed both the NEO-PI-R and AB5C-IPIP They were

re-cruited by mail from lists of homeowners and agreed to complete

questionnaires, delivered by mail, for pay, over a period of many

years, beginning in 1994 The sample spanned all levels of

edu-cational attainment, with an average of 2 years of postsecondary

schooling Most participants identified as White (97%), and 1% or

less (for each category) identified as Hispanic, Asian American,

Native American, or did not report their ethnicity

Measures. The NEO-PI-R (Costa & McCrae, 1992b) contains

240 5-point Likert scale items and breaks each of the Big Five

down into six facets, each assessed by eight items Costa and

McCrae (1992b) list internal reliabilities for the facet scales

rang-ing from 62 to 82 Similar reliabilities were obtained in the

present sample The NEO-PI-R was administered to the ESCS in

the summer of 1994

The AB5C-IPIP (Goldberg, 1999) contains 485 5-point Likert

scale items and breaks each of the Big Five down into nine facets,

each assessed by 9 –13 items The 45 AB5C-IPIP facet scales were

created on the basis of the content of the lexical AB5C facets,

using the IPIP, which was administered to the ESCS between 1994

and 1996 Internal reliabilities range from 67 to 90.2

Analysis. Factor analyses were performed using principal-axis

factoring (also known as common factor analysis), with direct

oblimin rotation (⌬ ⫽ 0) to allow correlated factors For the factor

analyses within each domain, the number of factors to extract was

determined using Velicer’s minimum average partial (MAP) test

(O’Connor, 2000) In the MAP test, a complete principal

nents analysis is performed, after which the first principal

compo-nent is partialed out of the correlations among the variables, and

the average squared partial correlation is noted This procedure is

repeated using the first two principal components, then the first

three, and so on The number of factors to extract is the number of

components that resulted in the minimum average squared partial

correlation This is the number of factors that are related to

systematic variance in the original correlation matrix

The MAP test’s ability to identify only those factors that are related to systematic variance in the matrix is particularly useful in the present context because of the likelihood of redundancy among facets across the two inventories Two facet scales measuring the same construct and thus having very similar content might be correlated strongly enough to split off and form their own factor Such a factor would simply reiterate the existence of that specific facet and would be uninformative for the purpose of investigating

a level of organization between facets and domains The MAP test would be unlikely to identify such a small factor

Results

Before factoring the 15 facets within each domain separately,

we examined the factor structure of all 75 facets together to make sure they conformed to the Big Five structure, as expected.3The first 10 eigenvalues were 15.19, 10.47, 8.57, 6.23, 5.02, 1.74, 1.54, 1.42, 1.34, 1.24 After extracting and rotating five factors, all facets had their highest loading on the expected factor, except for Trust and Assertiveness from the NEO-PI-R and Reflection from the AB5C-IPIP, and these three had strong secondary loadings on the expected factor (Trust loaded at ⫺.52 on Neuroticism and at 43 on Agreeableness; Assertiveness loaded at 56 on Conscien-tiousness and at 50 on Extraversion; Reflection loaded at 51 on Agreeableness and at 50 on Openness/Intellect.) Thus, there ap-pears to be no reason to exclude any facets from the analysis of individual Big Five domains

For the 15 facets within each Big Five domain, mere examina-tion of the eigenvalues (see Table 1) might suggest only one large factor Nonetheless, the MAP test indicated two factors in each domain (see Table 2), with one exception, Extraversion, for which three factors were indicated However, when Excitement Seeking was excluded from the MAP test for Extraversion, only two factors were indicated (see Table 2) A factor created by the presence of

a single-facet scale seems unlikely to be sufficiently broad to represent a meaningful factor at the level between facets and domains Furthermore, Excitement Seeking is the best marker of impulsivity within Extraversion (Whiteside & Lynam, 2001), and impulsivity is likely to be relatively peripheral to Extraversion (Depue & Collins, 1999) We therefore extracted two factors from each of the Big Five domains We retained Excitement Seeking in the analysis of Extraversion facets in order to examine its loadings

in the two-factor solution (Excluding it did not noticeably change the solution or scores for this factor, which were correlated at 999, with the factor scores from the analysis reported here.)

Table 3 shows the factor loadings and correlations within each domain and provides labels that attempt to capture the essence of each factor An additional column at the left in Table 3 contains codes for secondary loadings on the basis of the AB5C lexical model (Goldberg, 1999; Hofstee et al., 1992; Johnson, 1994) Note that these secondary loadings were not derived from the present factor analyses, but from calculations of the AB5C model in other samples These codes are discussed below

2The AB5C-IPIP is publicly available at http://ipip.ori.org/

3Descriptive statistics, the correlation matrix for all 75 facets, and the factor loadings for the five-factor solution are available from Colin G DeYoung upon request

Each of the Big Five was found to contain two distinct, though

correlated, factors underlying the variance shared among 15 facet

scales Before attempting to interpret the content of these factors,

we asked ourselves whether the presence of exactly two factors in

all five domains might simply be an artifact stemming from the

manner in which the facets of the AB5C-IPIP were constructed

Remember that 40 of the 45 AB5C facets are defined by a positive

loading on their primary domain and either a positive or negative

secondary loading on one other domain (the other five facets are

defined by descriptors loading exclusively on their primary

do-main and are thus factor-pure) All of the positive poles of the Big

Five are socially desirable, whereas all of the negative poles are

socially undesirable (Neuroticism is reversed in the AB5C and

labeled Emotional Stability), which might lead to two-factor

solu-tions in which traits with desirable and undesirable secondary

loadings clustered separately

In other words, our findings could be nothing but a social

desirability artifact In order to evaluate this possibility, we

exam-ined the division of positive and negative secondary loadings

(noted in Table 3) among the two factors for each domain Johnson

(1994) calculated the AB5C primary and secondary loadings for

the NEO-PI-R facets, so we were able to assign all 75 facets’

secondary loadings on the basis of the AB5C model (Note that the

codes for NEO Neuroticism facets are reversed in sign in order to

maintain the association between positive secondary loadings and

social desirability across all scales.)

What is immediately clear is that the facets do not consistently

split according to the social desirability of their secondary

load-ings All but 2 of the 10 factors are marked by facets with both

positive and negative secondary loadings Of interest as well is

that, within Agreeableness, Neuroticism, and Openness/Intellect,

factor-pure facets serve as markers of both factors These findings

bolster our supposition that factors within the facets of each Big

Five domain are likely to represent substantive and meaningful

distinctions in content rather than mere artifacts (Of course, Jang

et al.’s, 2002, finding of two genetic factors within each of the Big Five offers additional support for this position, as genes cannot be affected by social desirability.)

Each of the Big Five can thus be said to have two aspects, representing related but separable trait dimensions How should these dimensions be interpreted and labeled? The task is most straightforward for Openness/Intellect The long-running debate over the interpretation of this domain has left us with obvious choices to represent factors marked by facets like Quickness, Ingenuity, and Ideas, on the one hand, and Aesthetics, Imagination,

and Fantasy on the other: Intellect and Openness As other

re-searchers have noted, it appears that the two sides of this debate were simply focusing on different aspects of the larger domain (DeYoung et al., 2005; Johnson, 1994; Saucier, 1992) The factors that emerged here do not merely reflect the agendas of the authors

of our two instruments, who happen to fall on opposite sides of the Openness/Intellect debate, because two AB5C-IPIP facets are good markers of Openness and one NEO-PI-R facet is a good marker of Intellect

The two aspects of Extraversion are consistent with distinctions drawn in the literature between agency or dominance and

socia-bility We suggest Assertiveness and Enthusiasm as labels for these

two aspects of Extraversion While Assertiveness should be rela-tively uncontroversial as a compromise between the more general and abstract idea of agency and the more socially specific idea of dominance, Enthusiasm probably needs more thorough

justifica-tion Sociability is problematic as a descriptor of this aspect of

Extraversion because it focuses exclusively on the manner in which this trait is manifested socially, ignoring the crucial affec-tive component Along with Gregariousness and Friendliness, the Positive Emotions facet is a strong marker of this factor, and conceptions of Extraversion often focus on the tendency to expe-rience positive emotions associated with anticipation or enjoyment

of reward (Depue & Collins, 1999; Lucas, Diener, Grob, Suh, &

Table 1

Eigenvalues for Factor Analysis of 15 Facets in Each Big Five

Domain

Note N ⫽481 Principal-axis factoring N ⫽ Neuroticism; A ⫽

Agree-ableness; C ⫽ Conscientiousness; E ⫽ Extraversion; O ⫽ Openness/

Intellect

Table 2

MAP Test for Facets in Each Big Five Domain

3 040 031 039 .0448(.052) 032

Note. Numbers in parentheses are based on calculations excluding NEO Excitement Seeking The lowest average square partial correlation for each domain is in bold N ⫽ Neuroticism; A ⫽ Agreeableness; C ⫽ Consci-entiousness; E ⫽ Extraversion; O ⫽ Openness/Intellect

Table 3

Two-Factor Solutions for Each Big Five Domain

Secondary

loading

code Facet and instrument

Neuroticism Secondary

loading code Facet and instrument

Extraversion

Agreeableness Compassion Politeness

Conscientiousness Industriousness Orderliness

I⫹ Achievement striving

(NEO)

Note N ⫽481 Principal-axis factoring with direct oblimin rotation AB5C ⫽ Abridged Big Five Circumplex Scales from the International Personality Item Pool; I ⫽ Extraversion; II ⫽ Agreeableness; III ⫽ Conscientiousness; IV ⫽ Emotional Stability; V ⫽ Openness/Intellect; P ⫽ factor-pure; see text for discussion of these codes

Openness/Intellect Intellect Openness

Shao, 2000; Watson & Clark, 1997) Social interaction is often

rewarding, which appears to provide the motivation for the

socia-bility associated with Extraversion (Lucas & Diener, 2001)

En-thusiasm is a good label for this factor because it is broad enough

to describe both positive emotion and outgoing friendliness or

sociability John (1990) demonstrated that enthusiastic is an

ex-cellent descriptor of prototypical Extraversion

Our two Conscientiousness factors are nearly identical to factors

found in the same sample by Roberts et al (2005), in their analysis

of scales conceptually related to Conscientiousness from seven

different instruments.4We have therefore elected to use labels very

similar to theirs, Industriousness and Orderliness Orderliness

seems preferable to their term “Order” because the former

de-scribes a tendency of the individual, whereas the latter dede-scribes an

outcome of behavior or some other ordering process

The two aspects of Agreeableness appear to distinguish between

compassionate emotional affiliation with others (e.g., Warmth,

Sym-pathy, Tenderness) and a more reasoned (or at least cognitively

influenced) consideration of and respect for others’ needs and desires

(e.g., Cooperation, Compliance, Straightforwardness) We therefore

suggest Compassion and Politeness as labels for these factors

Polite-ness appears similar to Ashton and Lee’s (2005; Ashton et al., 2004)

Honesty-Humility factor, as both are marked by the NEO-PI-R facets

Straightforwardness and Modesty Given that AB5C-IPIP facets like

Morality and Compliance also mark this factor, Ashton and Lee’s

(2005) assertion that the NEO-PI-R is unlike other Big Five measures,

in containing content that could be included in their Honesty-Humility

factor, may be unfounded

The two factors within Neuroticism, which we labeled Volatility

and Withdrawal, are consistent not only with the lexical division

noted by Saucier and Goldberg (2001) between irritability and

anxi-ety/fearfulness but also with a tradition that distinguishes between

externalizing and internalizing problems (Achenbach & Edelbrock,

1978, 1984; Krueger, 1999) Facets like Stability (reversed), Angry

Hostility, and Impulsiveness imply problems of disinhibition, leading

to the outward expression of negative affect, whereas facets like

Depression, Vulnerability, and Anxiety imply problems of inhibition,

negative affect directed inward We chose the label Volatility because

it seems broad enough to encompass emotional lability, irritability or

anger, and difficulty controlling emotional impulses The second

factor appears to reflect susceptibility to a class of negative affect that

has commonly been described as withdrawal (Davidson, 2001) The

label Happiness, for the facet of the AB5C-IPIP that (reversed in sign)

is the strongest marker of the Withdrawal factor, is potentially

mis-leading because its items emphasize negative affect (“Seldom feel

blue,” “Feel threatened easily”) rather than positive affect

Choosing suitable labels for each factor obviously depends

heavily on interpretation of the factors’ content, which can be

difficult when based merely on facet labels Furthermore,

inter-preting factors that are fairly strongly correlated poses an

addi-tional challenge, as many facets load strongly on both factors We

therefore defer further justification of our interpretations until

Study 2, in which we examine individual items that best mark each

of the 10 aspect factors

Study 2 The IPIP contains over 2,000 public domain items that have

been administered to the ESCS, on which we performed our

analysis in Study 1 It is thus uniquely well suited to the empirical characterization of factor content at the item level We examined correlations between scores for the 10 aspect factors presented in Table 3 and every IPIP item

In addition to allowing more precise characterization of the aspect factors, this undertaking had the advantage of allowing the creation of an instrument to measure the 10 aspects of the Big Five Such an instrument would allow the aspects to be assessed in other samples without having to administer two very long questionnaires and perform multiple factor analyses Given that the NEO-PI-R is widely used, another strategy, especially for existing data, would

be to use the factor loadings presented in Table 3 to identify NEO facets or combinations of facets that are good markers for each aspect One limitation of this strategy, however, is that no good markers for Compassion appear in the NEO-PI-R Two of the NEO-PI-R Agreeableness facets (Altruism and Tender-Mindedness) load strongly on Compassion, but they load almost equally on Politeness They are good markers, therefore, of Agree-ableness as a whole, but they cannot discriminate Compassion from Politeness Additionally, administration of the NEO-PI-R is costly and time-consuming, and a shorter instrument designed specifically to assess the 10 aspects of the Big Five might be preferable in many situations We therefore took advantage of the IPIP to develop such an instrument, the Big Five Aspect Scales (BFAS)

4A question raised by differences between Roberts et al.’s (2005) results and ours is why they found six factors, whereas we found only two The statistical answer is that we used only two of the seven instruments that they used, and, even in their study, all but two of the scales from these two instruments fell within two factors Of course, the real question is whether the NEO-PI-R and AB5C-IPIP neglect some facets of Conscientiousness

We suspect not Rather, it appears that Roberts et al.’s (2005) additional factors are best viewed as compound traits, stemming from the conjunction

of Conscientiousness with other traits, rather than as aspects or facets of Conscientiousness itself Roberts et al.’s Self-Control factor is marked by two scales from the Hogan Personality Inventory (HPI; Hogan & Hogan, 1992), Impulse Control and Not Spontaneous, that have their primary loading on Extraversion rather than on Conscientiousness in the AB5C model (Johnson, 1994) Similarly, their Virtue factor is marked by two HPI scales, Moralistic and Virtuous, that do not have their primary or secondary AB5C loadings on Conscientiousness (Johnson, 1994) (This situation highlights one pitfall of personality research: The fact that a scale has been conceptually located in one of the Big Five domains may not be the best guide to determine whether it is statistically located in that domain.) Traditionalism and Responsibility also seem likely to be compound traits (though AB5C codes have not been calculated for all of the scales that mark them) Traditionalism appears to indicate conformity with moral norms, which we (DeYoung et al., 2002) have demonstrated can best be located within the Big Five hierarchy at the metatrait level, as a compound trait resulting from the combination of high Stability (the shared variance

of Emotional Stability, Conscientiousness, and Agreeableness) and low Plasticity (the shared variance of Extraversion and Openness/Intellect) Roberts et al described Responsibility as reflecting enjoyment of cooper-ation and being of service to others, which suggests Agreeableness as much

if not more than Conscientiousness We conclude that additional Conscientiousness-related factors beyond Industriousness and Orderliness

do not appear best described as lower order traits within the domain of Conscientiousness, though they are interesting constructs in their own right and may be useful in the prediction of behavior

Following selection of items that were good markers of each

aspect in the ESCS, these items were administered to a large

university sample Once the final items were selected on the basis

of their psychometric properties in the university sample, we were

able to examine the reliability and validity of the instrument in

both samples

Method

Initial item selection. Factor scores for each of the 10 factors

presented in Table 3 were calculated using the regression method

These scores were then correlated with all of the IPIP items As an

initial item pool, we chose 15 items showing the highest

correla-tions with each factor, excluding those that seemed overly

redun-dant and making sure to include roughly equal numbers of

posi-tively and negaposi-tively keyed items In order to provide adequate

discrimination between the two aspects in each domain, and to

prevent excessive cross-loadings on other domains, we excluded

items that showed a correlation with another factor within 10 of

the primary correlation For example, if the strongest correlation

for a particular item was 58 with Compassion, then we would

exclude it if its correlation with Politeness or any of the other eight

aspect factors was 48 or greater.5

Having selected 150 IPIP items to mark the 10 aspects, we

administered them to a large undergraduate sample, intending to

choose 10 items to measure each aspect, based on their

psycho-metric properties in the new sample, for a total of 100 items Prior

to administration, we changed the wording for three of the

nega-tively keyed items selected for Politeness, in order to reverse their

keying direction, because only two positively keyed items in the

IPIP met our selection criteria for this aspect (see Table 4)

Additionally, we added a new item, “Am not a very enthusiastic

person,” to test our hypothesis that Enthusiasm is a good label for

this aspect of Extraversion

Participants and measures. Participants were 480

undergradu-ates in southern Ontario (299 women and 180 men; 1 with no gender

reported), enrolled at the University of Toronto, Toronto, Ontario,

Canada, or the University of Waterloo, Waterloo, Ontario, Canada

They ranged in age from 17 to 61 years (M ⫽ 19.32, SD ⫽ 3.33) and

came from diverse ethnic backgrounds (45% White; 34% East Asian;

9% South Asian; 3% Black; 3% Middle Eastern; 1% Hispanic; 5%

unknown) All participants received course credit for completing the

study The potential BFAS items and the Big Five Inventory (BFI;

John & Srivastava, 1999), were completed via the Web, using Likert

scales ranging from 1 to 5 The BFI, which was completed by 472

participants, is an excellent short measure of the Big Five and thus

makes a good benchmark against which to validate new Big Five

scales (Additionally, 423 of our ESCS participants also completed

the BFI, allowing comparison across samples.)

Approximately 1 month following their completion of the study,

participants were contacted by e-mail and asked to complete the

BFAS items again via the Web in order to obtain an index of

test–retest reliability Ninety participants completed the retest, and

the average number of days between first and second completion

of the BFAS was 38.44 (SD ⫽ 10.71).

Results

Final item selection. Principal-axis factoring with direct

ob-limin rotation (⌬ ⫽ 0) was used to extract two factors from the

items in each of the Big Five domains In order to reduce col-linearity in the final scales, items were included only if their loading on the intended aspect factor was at least 10 greater than

on the other aspect factor This criterion was used to exclude 20 items, but it was relaxed for 5 other items in order to maintain balanced keying No scale was allowed a ratio of positively to negatively keyed items (or vice versa) greater than 6/4 Addition-ally, a five-factor solution was extracted from all items across all five domains, and items were excluded if they did not have their highest loading on the intended Big Five domain; 14 items were excluded by this criterion.6

Table 4 shows the 10 final items for each of the 10 scales Right columns in Table 4 show the correlation of each item with the relevant factor score from the ESCS in Study 1 and the factor loading of each item on the relevant aspect factor in the university sample from Study 2 Items were averaged (with appropriate reversals) to create scale scores for each aspect, and these scores were averaged across the two aspects in each domain to create Big Five domain scores Thus, in addition to 10-item scales for the 10 aspects, the BFAS includes 20-item scales for the Big Five

Reliability and validity of the BFAS. Table 5 provides descrip-tive statistics for the BFAS, including Cronbach’s alpha for the

ESCS (M ⫽ 0.83, SD ⫽ 0.03), the initial university sample (M ⫽ 0.81, SD ⫽ 0.05), and the retest university sample (M ⫽ 0.83,

SD ⫽ 0.05) (There were no significant differences in BFI or BFAS scores between those who completed the retest and those who did not, nor did scores change significantly from test to retest.) Correlations between scale scores and factor scores from

Study 1 are given for the ESCS (M ⫽ 0.89, SD ⫽ 0.02), and test–retest correlations are given for the university sample (M ⫽ 0.81, SD ⫽ 0.04) Table 6 contains correlations between all BFI

and BFAS scales Correlations between the same Big Five do-mains across scales (in bold italics) were high; when corrected for

attenuation, based on reliability, they ranged from 85 to 96 (M ⫽ 0.90, SD ⫽ 0.05) for the university sample and from 72 to 91 (M ⫽ 0.84, SD ⫽ 0.07) for the ESCS Table 6 also reveals that

patterns of correlation among the Big Five within each instrument

5One effect of this selection procedure was to exclude items that appear most central to each of the Big Five domains because they are related strongly but almost equally to both aspects These items are potentially informative conceptually For example, the item “Have a vivid imagina-tion” was associated almost equally with Intellect and Openness, support-ing Saucier’s (1992) suggestion of Imagination as an alternative label for the Openness/Intellect domain The argument that unconventionality is also important to this domain (de Raad, Perugini, Hrebickova, & Szarota, 1998) finds some support in the excluded item “Like to be viewed as proper and conventional.” Other insights from these excluded items include the fact that the talkativeness associated with Extraversion is characteristic

of both Enthusiasm and Assertiveness (“Usually like to talk a lot”; “Have little to say”) and that susceptibility to stress and negative emotions appears common to both Volatility and Withdrawal (“Get stressed out easily”; “Am often in a bad mood”)

6For example, the item “Tend to vote for liberal political candidates” was a clear marker of Openness in the ESCS but had its strongest load-ing—negatively— on Conscientiousness in the university sample This finding is not particularly surprising, as Goldberg and Rosolack (1994) found that conservatives were low in Openness/Intellect but high in Con-scientiousness, but it does suggest that this item is not a good specific marker of Openness

Table 4

The Big Five Aspect Scales

Scale

rwith factor score (ESCS)

Factor loading (University) Neuroticism

Volatility

Am a person whose moods go up and down

easily

Withdrawal

Agreeableness

Compassion

Politeness

Conscientiousness

Industriousness

(table continues)

Table 4 (continued )

Scale

rwith factor score (ESCS)

Factor loading (University) Orderliness

Extraversion

Enthusiasm

Assertiveness

Openness/Intellect

Intellect

Openness

Seldom notice the emotional aspects of paintings and pictures (R) ⫺.60 ⫺.47

Note. Items from all 10 scales should be interspersed for administration, and 5-point Likert scales should be used for responses (R) indicates items to be reverse scored; ESCS ⫽ Eugene-Springfield community sample

aThese items were keyed in the opposite direction for the ESCS

bThis item is new; it was not included in the International Personality Item Pool or administered to the ESCS

(in bold) are similar, offering further support for similarity of

measurement across instruments

In the ESCS, we were additionally able to validate the BFAS

against NEO-PI-R domain scores and Saucier’s (1994)

Mini-Markers, a well-validated adjective marker set for the lexical Big

Five, which participants completed at the same time as the BFI

(see Table 7) High correlations between the same Big Five

do-mains across scales (in bold) provide an additional demonstration

that the BFAS is measuring the standard Big Five When corrected

for attenuation, these correlations ranged from 80 to 92 (M ⫽

0.88, SD ⫽ 0.05) for the NEO-PI-R and from 80 to 85 (M ⫽ 0.82,

SD ⫽0.02) for the Mini-Markers

Discriminant validity and an example of suppression. Given

the fairly strong correlations between the two aspect factors in

each domain, one important question is: To what degree do the two

aspects of each domain possess discriminant validity? If the two

aspects within each Big Five domain are indeed distinct traits, then

they should not show overly similar patterns of correlation with

other variables Table 6 confirms that they do not, for all five

aspect pairs The differential associations of the aspect pairs of

Extraversion and Agreeableness provide one clear example:

Whereas Assertiveness is negatively correlated with Politeness,

Enthusiasm is positively correlated with Politeness

Because each pair of aspects is positively correlated, assessing

discriminant validity can be more complicated than simply looking

for divergent patterns of zero-order correlations Being positively

correlated and presumably sharing some of the same sources, the

two aspects in each domain should predict many variables

simi-larly Furthermore, whenever they do not predict some variable

similarly, they may act as suppressors on each other When two

positively correlated variables are related to a third variable in

opposite directions, one or both of their associations with the third

variable may be suppressed (Paulhus, Robins, Trzesniewski, &

Tracy, 2004) Multiple regression or partial correlation may then

be necessary to control for the positive association between the first two variables in order to examine the unique associations of their nonshared variance with the third variable (Although the correlations between aspects are fairly strong, none of them reach

the threshold [r ⬎ 9] at which multicollinearity typically becomes

a problem for such analyses; Tabachnick & Fidell, 2001)

As one example of suppression, consider the associations of the aspects of Conscientiousness with BFI Neuroticism (see Table 6)

In previous research, the negative correlation between Conscien-tiousness and Neuroticism has proved to be one of the most robust cross-domain correlations among the Big Five (Mount, Barrick, Scullen, & Rounds, 2005) Using the BFAS, however, one can see that this correlation holds only for Industriousness Orderliness is almost uncorrelated with Neuroticism Not only that, but when one controls for Industriousness, Orderliness is significantly positively correlated with Neuroticism, in both the university sample and the

ESCS (University: partial r ⫽ 24, p ⬍ 01; ESCS: partial r ⫽ 20,

p ⬍.01) Thus, the negative association between Industriousness and Neuroticism was suppressing a positive association between Orderliness and Neuroticism

Correlations among the aspects. Patterns of correlation among the aspect-level traits (bottom right corner of Table 6) are more varied than correlations among domains, and stronger cross-domain correlations appear at the aspect level than at the Big Five level In several cases, correlations between two aspects across two domains are at least as strong as correlations between the two aspects within each of those two domains This is true of the correlations between Intellect and Industriousness and between Intellect and Assertiveness (In fact, Intellect, Industriousness, and Assertiveness form a cluster of related scales from three different domains.) Could this finding be a product of our final item selec-tion procedure, which intenselec-tionally reduced correlaselec-tions between aspects within the same domain, by choosing items that discrim-inated well between the two aspects? This explanation seems

Table 5

Descriptive Statistics for the BFAS in Two Samples

Factor

Note. BFAS ⫽ Big Five Aspect Scales; ESCS ⫽ Eugene-Springfield Community Sample; ␣1⫽ internal reliability in original sample (N ⫽ 480); ␣2⫽

internal reliability in retest sample (N ⫽ 90).

aCorrelation with factor scores from Study 1, Table 3

bTest–retest correlation